Select one discovered HTP session instead of ['HTP*']. Use its exact runtime name; preserve CPU fallback and handle existing saved selections.
Evidence: PR #901 benchmark report.
| Myron · Qwen3 1.7B Q4_K_M |
Six sessions, pipeline off |
One session, stock |
| Prefill tok/s |
209.98 |
211.51 |
| Decode tok/s |
19.25 |
20.22 |
| Allocated MiB |
1,854.15 |
1,764.11 |
| Scope |
Detail |
| Motivation |
Avoid rc.3 multi-session pipeline memory amplification; no throughput loss measured here |
| S23 |
Already exposes one session; selection stays equivalent |
| Evidence limit |
Native diagnostic, flash ON; proposed APK change not yet tested |
| Acceptance |
Validate APK on both phones, UI flash OFF and bench flash ON, including previously failing large models; check memory, throughput and generation |
Select one discovered HTP session instead of
['HTP*']. Use its exact runtime name; preserve CPU fallback and handle existing saved selections.Evidence: PR #901 benchmark report.